|
|
|
|
|
|
by kennyfrc
1310 days ago
|
|
|
This looks great! I’ve played a couple of games in the Open Sicilian and the French Tarrasch and there wasn’t a moment that I felt like I was playing against a computer. It plays some cheezy lines which I would expect from a lichess player. On the user side - I think the review area would benefit from principle-based analysis similar to what is seen in chess.com. I am guessing that for it to happen, there needs to be a separate analysis engine that can detect positional motifs. I think chess.com did a decent job with that (better and more actionable than decodechess in my opinion). On the tech side - I’m curious how this project differs from existing neural nets trained for lc0 such as bad gyal and maia? |
|
|
On the tech side, what Maia is doing I think is similar to our NN component. It turns out to get realistic play (include randomness, avoid too many one-move blunders, same level of tactical play as positional play, realistic endgame play, etc.) you have to do quite a lot on top of that – much more than I initially thought. I have tested the AI over the last year with friends of different strengths to try to make the gameplay more realistic and engaging. I also think the automatic adaptation to your level of play, and predicting the rating, is unique. Bad gyal AFAIK uses a different training method where Stockfish evaluations is involved, but I haven't really tested it enough to know how it stands vs. Maia.